An immune cell image data classification and recognition system

Abstract

The application relates to the technical field of image data processing, in particular to an immune cell image data classification and recognition system, which comprises a space-time registration module, a dynamic trajectory analysis module, a membrane surface deformation detection module and a subtype classification decision module.In the application, the coordinate system is aligned through affine transformation, the registration precision is improved in combination with a gray variance threshold screening mechanism, three-frame sliding window is used to calculate a gray gravity center displacement vector, the direction convergence is verified through cosine similarity, the trajectory continuity capturing capability is enhanced, noise interference pseudo-trajectory is filtered, a curvature-area coupling model is used to quantize membrane boundary gradient resistance, three-point extreme detection is used to position a deformation region, the limitation of edge detection geometric representation is broken through, a support vector machine is used to analyze a trajectory consistency coefficient, a dynamic threshold segmentation is used to quantize resistance mutation, a multi-dimensional nonlinear discrimination boundary is constructed, the subtype discrimination specificity and sensitivity are improved, space-time registration, trajectory verification, mechanical analysis and feature fusion progressive processing are carried out, and behavior feature full-dimensional analysis is realized.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to an immune cell image data classification and recognition system. Background Technology

[0002] Image data processing technology encompasses the acquisition, analysis, interpretation, and computation of image information. The core of this technology lies in converting raw images into analyzable data through digital means, and further performing processing operations such as recognition, classification, matching, and structure extraction. Its overall technology covers image preprocessing, image feature extraction, image recognition, classification, and its fusion analysis with other data types. Image data processing is widely used in various fields such as medical, industrial automation, security, and transportation, with particular emphasis on the precise analysis and interpretation of biomedical images of cells, tissues, and organs in the medical image processing field.

[0003] The immune cell image data classification and recognition system refers to a classification and recognition method and device established for immune cell image data. This patent primarily concerns the localization, extraction, labeling, and recognition process of immune cells in images. It covers methods such as image grayscale distribution, edge contour detection, and color channel analysis to segment cell regions; constructing a feature extraction structure based on a convolutional neural network to obtain local and global image features; subsequently, using a multilayer perceptron in a supervised learning model for classification; and completing the correspondence between image data and cell types using a label mapping method. The entire process completes the recognition output in an end-to-end manner after standardized image input.

[0004] Existing technologies, based on convolutional neural networks to extract static image features, lack mechanisms for capturing temporal dynamic changes, leading to accumulated frame-breaking errors in motion trajectory analysis. Multilayer perceptron classification models rely on single-frame image features, failing to integrate the correlation between membrane surface deformation and cytoplasmic motion trajectories, resulting in high misclassification rates in live cell dynamic analysis scenarios. Traditional segmentation methods locate cell regions through color channels or edge contour detection, but lack accuracy in identifying cell boundaries with low contrast or overlapping morphologies, and fail to establish quantitative models of deformation and mechanical properties. End-to-end systems employ linear feature transfer mechanisms, lacking trajectory direction convergence verification and deformation abrupt change amplitude quantification, making it difficult to distinguish subtypes with similar appearance features. For example, in T cell subtype identification, existing technologies cannot analyze differences in chemotactic motion trajectories and changes in membrane surface mechanical properties, resulting in insufficient matching between classification results and cell functional characteristics. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an immune cell image data classification and recognition system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an immune cell image data classification and recognition system comprising:

[0007] The spatiotemporal registration module is used to align the coordinate system of multiple time-series images through an affine transformation algorithm based on a dynamic threshold of gray-level variance in spatiotemporal correlation, filter cytoplasmic candidate regions based on dynamically adjusted gray-level variance thresholds, and transfer the coordinate set of the cytoplasmic candidate regions to the dynamic trajectory analysis module and transfer the membrane boundary gradient extreme points to the membrane surface deformation detection module.

[0008] The dynamic trajectory analysis module is used to call the optical flow equation parameter optimization method based on spatiotemporal constraints to analyze the coordinate set of the cytoplasmic candidate region, calculate the gray-level centroid displacement vector through a three-frame sliding window, verify the directional convergence using cosine similarity, and pass the effective trajectory vector to the subtype classification decision module.

[0009] The membrane surface deformation detection module is used to calculate the resistance index of the membrane boundary gradient extreme point using a curvature-area coupling model that includes a curvature change rate weighting coefficient α and an area change rate weighting coefficient β. It identifies the coordinates of the deformation region through a three-point extreme value detection method and transmits the coordinates of the deformation region to the subtype classification decision module.

[0010] As a further embodiment of the present invention, the cytoplasmic candidate region coordinate set specifically includes the candidate region center coordinates and region boundary coordinates; the membrane boundary gradient extreme points include extreme point coordinates and gradient magnitude; the effective trajectory vector includes trajectory direction, displacement magnitude, and convergence coefficient; and the deformation region coordinates specifically include the deformation region center position and region boundary range.

[0011] As a further aspect of the present invention, the dynamically adjusted grayscale variance threshold is determined based on the weighted average of the standard deviations of the grayscale distributions of three adjacent frames.

[0012] The regularization parameter λ in the optical flow equation parameter optimization method has a value range of 0.1-0.3, which is used to balance the weight ratio of spatial gradient and temporal gradient.

[0013] The curvature-area coupling model satisfies ,in Indicates the rate of change of curvature in the current frame. This represents the rate of change of the current frame area, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.4.

[0014] As a further aspect of the present invention, the spatiotemporal registration module includes:

[0015] The coordinate registration submodule is used to acquire multi-frame temporal image data, establish a spatial transformation model between adjacent frames, calculate the difference in translation, rotation angle, and scaling ratio, linearly combine the difference parameters using an affine transformation matrix, iteratively optimize the matrix coefficients using the weighted least squares method, and generate a registration parameter set.

[0016] The cytoplasmic screening submodule is used to call the registration parameter set to perform image coordinate transformation, extract the temporal grayscale sequence of multiple pixels, calculate the difference between the sequence variance value and the activity threshold, mark the pixel coordinates with a difference greater than zero in the continuous region, perform 8-neighborhood connectivity detection to merge fragmented regions, and establish a cytoplasmic candidate region coordinate set.

[0017] The membrane boundary extraction submodule is used to call the registration parameter set to transform the image coordinate system, extract the gradient sequence along the normal direction of the membrane structure edge, calculate the rate of change of gradient magnitude of adjacent pixels after normalization, detect abrupt change points where the rate of change exceeds the mean plus twice the standard deviation, record the coordinates of abrupt change points to form a spatial distribution set, and obtain the extreme points of the membrane boundary gradient.

[0018] In the weighted least squares method, the spatiotemporal correlation weight factor γ is dynamically adjusted according to the area ratio of the overlapping region of the image.

[0019] The normalization process uses the maximum-minimum standardization method to map the gradient magnitude to the [0,1] interval.

[0020] As a further aspect of the present invention, the dynamic trajectory analysis module includes:

[0021] The optical flow tracing submodule is used to call the coordinate set of the cytoplasmic candidate region, establish the gray-level gradient matrix of adjacent frames, calculate the ratio of spatial gradient to temporal gradient in the optical flow equation, solve the displacement vector by the least squares method with regularization constraints, and perform median filtering on the vector using a three-frame sliding window to generate an initial trajectory vector set.

[0022] The direction verification submodule is used to extract the displacement vectors of three consecutive frames of the initial trajectory vector set, calculate the cosine value of the angle between adjacent vectors, compare the cosine value with the direction convergence threshold, mark the trajectory segments whose comparison results exceed the threshold for three consecutive times, and establish a candidate trajectory segment set.

[0023] The trajectory filtering submodule is used to statistically analyze the standard deviation of the displacement direction of multiple trajectory segments in the candidate trajectory segment set, calculate the difference between the standard deviation and the angle tolerance value, filter the trajectory segments with negative differences, integrate the displacement vector parameters of the trajectory segments that meet the conditions, and obtain the effective trajectory vector set.

[0024] The directional convergence threshold is determined by the directional consistency distribution of the effective trajectory vectors in the statistical training set, and the 85th percentile value of the cumulative distribution function of this distribution is taken.

[0025] As a further aspect of the present invention, the membrane surface deformation detection module includes:

[0026] The curvature gradient analysis submodule is used to obtain the extreme points of the membrane boundary gradient. It uses a curvature-area coupling model to simultaneously calculate the rate of curvature change and the rate of area change of the extreme points, establishes a curvature-area coupling parameter matrix, and obtains the extreme points of curvature gradient and the rate of area change through matrix eigenvalue decomposition.

[0027] The resistance index calculation submodule is used to call the curvature gradient extreme points and the area change rate, set the resistance coefficient weight in the curvature-area coupling model, perform a weighted superposition operation of the curvature change rate and the area change rate on the extreme points, and generate a resistance index distribution map corresponding to multiple extreme points.

[0028] The deformation region positioning submodule is used to compare the resistance index gradient of adjacent extreme points based on the resistance index distribution map using the three-point extreme value detection method, select the extreme point group whose gradient change exceeds a set threshold, calculate the center point of the deformation region through three-point coordinate interpolation, and generate the deformation region coordinate set.

[0029] The weighted superposition operation introduces a cell membrane elastic modulus correction factor η, the value of which is positively correlated with cell type.

[0030] The gradient change threshold is set to 60%-70% of the maximum gradient of the resistance index.

[0031] As a further aspect of the present invention, the system further includes:

[0032] The subtype classification decision module is used to receive the effective trajectory vector and the coordinates of the deformation region, analyze the trajectory direction consistency coefficient using a support vector machine based on the radial basis kernel function, quantify the resistance mutation amplitude by a dynamic threshold segmentation algorithm combined with a sliding window standard deviation feedback mechanism, generate cell subtype activation labels, and output the cell subtype activation labels to the storage unit.

[0033] As a further aspect of the present invention, the cell subtype activation label includes trajectory direction consistency coefficient, resistance mutation amplitude, and subtype classification result.

[0034] As a further aspect of the present invention, the initial threshold of the dynamic threshold segmentation algorithm is set as the moving average of the resistance mutation amplitude of the previous five frames.

[0035] As a further aspect of the present invention, the subtype classification decision module includes:

[0036] The trajectory consistency analysis submodule is used to obtain the effective trajectory vector and the coordinates of the deformation region, map the trajectory direction feature space using the support vector machine kernel function, calculate the vector dot product value between each trajectory point and the region coordinate axis, normalize the dot product value to obtain the direction angle cosine sequence, and apply sliding window mean filtering to the cosine sequence according to the exponential decay time series weight to generate the trajectory direction consistency coefficient.

[0037] The resistance mutation quantization submodule is used to divide the resistance mutation threshold interval based on the trajectory direction consistency coefficient using an adaptive Gaussian mixture model, calculate the second-order difference gradient value of the trajectory displacement in the X / Y axis of the deformation region, perform frame-by-frame difference operation on the absolute value of the gradient and the dynamic threshold, and calculate the cumulative amplitude sum of the points exceeding the threshold to generate the resistance mutation amplitude value.

[0038] The activation determination submodule is used to call the trajectory direction consistency coefficient and the resistance mutation amplitude value to establish a two-dimensional feature space coordinate system, construct a classification hyperplane using linear discriminant analysis, calculate the standardized Mahalanobis distance from the sample point to the decision boundary, divide the activation level interval according to the percentile of the distance distribution, and generate cell subtype activation labels.

[0039] The exponential decay factor δ is 0.85-0.95, corresponding to a time window length of 5-7 frames;

[0040] The dynamic threshold update mechanism uses a weighted combination of 90% of the threshold of the previous frame and 10% of the average gradient magnitude of the current frame.

[0041] The Mahalanobis distance calculation uses the inverse of the training set covariance matrix as the weight matrix.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In this invention, coordinate system alignment is achieved through affine transformation, combined with a gray-level variance thresholding mechanism to improve the spatial registration accuracy of cross-time series images and eliminate feature drift caused by cell displacement. The optical flow tracing algorithm uses a three-frame sliding window to calculate the gray-level centroid displacement vector, verifies directional convergence through cosine similarity, enhances the ability to capture the continuity of motion trajectories, and filters out pseudo-motion trajectories formed by noise interference. A curvature-area coupling model quantifies the resistance index of the membrane boundary gradient extremum points, and a three-point extremum detection method accurately locates the deformation region, overcoming the limitations of traditional edge detection in representing geometric features. Support vector machines analyze the trajectory direction consistency coefficient, and dynamic threshold segmentation quantifies the resistance mutation amplitude, establishing a nonlinear discrimination boundary for multi-dimensional feature fusion, improving the specificity and sensitivity of subtype discrimination. Spatiotemporal registration, motion trajectory verification, deformation mechanics analysis, and multi-modal feature fusion form a progressive processing logic, achieving full-dimensional analysis of dynamic behavioral features. Attached Figure Description

[0044] Figure 1 This is a system flowchart of the present invention;

[0045] Figure 2 This is a flowchart of the spatiotemporal registration module of the present invention;

[0046] Figure 3 This is a flowchart of the dynamic trajectory analysis module of the present invention;

[0047] Figure 4 This is a flowchart of the membrane surface deformation detection module of the present invention;

[0048] Figure 5 This is a flowchart of the subtype classification decision module of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0051] Example 1

[0052] Please see Figure 1 An immune cell image data classification and recognition system includes:

[0053] The spatiotemporal registration module is used to align the coordinate systems of multiple time-series images using an affine transformation algorithm based on a dynamic threshold of gray-level variance in spatiotemporal correlation. It filters candidate cytoplasmic regions based on a dynamically adjusted gray-level variance threshold, and transmits the coordinate set of candidate cytoplasmic regions to the dynamic trajectory analysis module. It also transmits the membrane boundary gradient extreme points to the membrane surface deformation detection module.

[0054] The dynamic trajectory analysis module is used to call the optical flow equation parameter optimization method based on spatiotemporal constraints to analyze the coordinate set of cytoplasmic candidate regions, calculate the gray-level centroid displacement vector through a three-frame sliding window, verify the directional convergence using cosine similarity, and pass the effective trajectory vector to the subtype classification decision module.

[0055] The membrane surface deformation detection module is used to calculate the resistance index of the membrane boundary gradient extreme point using a curvature-area coupling model that includes the curvature change rate weighting coefficient α and the area change rate weighting coefficient β. It identifies the coordinates of the deformation area through the three-point extreme value detection method and transmits the coordinates of the deformation area to the subtype classification decision module.

[0056] The subtype classification decision module receives the effective trajectory vector and the coordinates of the deformation region, uses a support vector machine based on the radial basis kernel function to analyze the trajectory direction consistency coefficient, quantifies the resistance mutation amplitude by combining a dynamic threshold segmentation algorithm with a sliding window standard deviation feedback mechanism, generates cell subtype activation labels, and outputs the cell subtype activation labels to the storage unit.

[0057] The cytoplasmic candidate region coordinate set specifically includes the coordinates of the candidate region center and the region boundary coordinates. The membrane boundary gradient extreme points include the extreme point coordinates and gradient magnitude. The effective trajectory vector includes the trajectory direction, displacement magnitude, and convergence coefficient. The deformation region coordinates specifically include the deformation region center position and the region boundary range. The cell subtype activation label includes the trajectory direction consistency coefficient, resistance mutation magnitude, and subtype classification result.

[0058] The dynamically adjusted grayscale variance threshold is determined based on the weighted average of the standard deviations of the grayscale distributions of three adjacent frames.

[0059] The regularization parameter λ in the optical flow equation parameter optimization method takes a value of 0.1-0.3 and is used to balance the weight ratio of spatial gradient and temporal gradient.

[0060] The curvature-area coupling model satisfies ,in Indicates the rate of change of curvature in the current frame. This represents the rate of change of the area in the current frame, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.4.

[0061] The initial threshold of the dynamic threshold segmentation algorithm is set as the moving average of the resistance mutation amplitude of the previous five frames.

[0062] Please see Figure 2 The spatiotemporal registration module includes:

[0063] The coordinate registration submodule is used to acquire multi-frame temporal image data, establish a spatial transformation model between adjacent frames, calculate the difference in translation, rotation angle, and scaling ratio, linearly combine the difference parameters using an affine transformation matrix, iteratively optimize the matrix coefficients using the weighted least squares method, and generate a registration parameter set.

[0064] The coordinate registration submodule begins by acquiring a set of time-series image data, specifically 10 consecutive frames of cell micrographs, each frame with a size of 512x512 pixels, denoted as . ,in Represents the frame number, from 1 to 10. Indicates the planar position of a pixel in the image, with a grayscale value ranging from 0 to 255. The processing flow unfolds on two adjacent frames, starting with the first frame. With the second frame For example, we first establish a spatial transformation model between them, here we choose the affine transformation model. This model is achieved by identifying and matching feature points between two frames of images, selecting 60 pairs of salient feature points. ,in yes The coordinates of the feature points in the image. yes The corresponding feature point coordinates are determined. Based on these matching point pairs, the translation difference is initially estimated as follows: 3.1 pixels in the x-direction and 2.4 pixels in the y-direction. , The difference in rotation angle is ,Right now And the scaling difference: the scaling ratio in the x-direction is 1.015, and the scaling ratio in the y-direction is 0.995, that is... .

[0065] Next, these difference parameters are linearly combined using an affine transformation matrix. Affine transformation matrix The form is To perform matrix calculations, the rotation angle is... It needs to be converted to radians, which is calculated as follows: Radius. This radian value and other parameters... , , , Substituting the values, we obtain the initial affine transformation matrix. :

[0066] ;

[0067] ;

[0068] Subsequently, the six free parameters of the matrix were analyzed using the weighted least squares method. Iterative optimization is performed. The objective function for optimization is defined as minimizing the transformation error of all selected pixels: ,in These are the pixels in the first frame. Its grayscale value, To transform the matrix The pixel grayscale value mapped to the corresponding position in the second frame. Weight Used to adjust the contribution of different pixels in the optimization process. Spatiotemporal correlation weighting factor. The confidence level used in this stage to adjust the overall optimization is dynamically adjusted based on the proportion of the overlapping area of ​​the image. The specific process of this adjustment is: estimating the area of ​​the overlapping region between the two transformed frames. For a 512x512 pixel image, the total area is... Square pixels. If the calculated area of ​​the overlapping region is 245760 square pixels, then the percentage of the overlapping region area is... At this point, the spatiotemporal correlation weighting factor That is, it is set to 0.9375. The value will be used as the global error term in the multiplier-adjusted weighted least squares method, meaning the objective function becomes... ,in The weights represent the values ​​of other factors. The matrix coefficients are repeatedly adjusted using iterative algorithms such as the Levenberg-Marquardt algorithm until the change in the objective function value is less than a preset convergence threshold (set to 0). ) or reach the maximum number of iterations (set to 150). This process applies to all adjacent frame pairs ( , until Each step is executed once, thereby generating a registration parameter set containing 9 optimized affine transformation matrices.

[0069] The cytoplasmic screening submodule is used to call the registration parameter set to perform image coordinate transformation, extract the temporal grayscale sequence of multiple pixels, calculate the difference between the sequence variance and the activity threshold, mark the pixel coordinates with a difference greater than zero in the continuous region, perform 8-neighborhood connectivity detection to merge fragmented regions, and establish a set of coordinates for cytoplasmic candidate regions.

[0070] The cytoplasm screening submodule first calls the registration parameter set generated by the coordinate registration submodule. This parameter set contains the affine transformation matrices between adjacent frames, such as from frame 1 to frame 2. Frame 2 to Frame 3 And so on. These matrices are used to perform image coordinate transformations, the purpose of which is to transform all frames in the sequence ( The coordinate system of the first frame In the coordinate system. For the second frame. any pixel in Its corresponding coordinates in the first frame coordinate system Through the Inverse the coordinates and then perform coordinate multiplication (i.e.) The calculation ensures that each pixel in the time series corresponds precisely in space.

[0071] Based on the aligned image sequence, a temporal grayscale sequence is extracted from multiple monitoring points (selecting 2000 predetermined pixel coordinates) distributed throughout the image. For a specific pixel coordinate... Its grayscale value sequence in 10 frames of images is recorded as follows An actual grayscale sequence is: Next, calculate the variance of this sequence. First, calculate the average gray value. Variance Then, the difference between this variance value and the preset activity threshold is calculated. The activity threshold is set with reference to the statistical characteristics of the temporal grayscale sequence of pixels in the image background region: 150 pixels in the image identified as background regions are selected, and the variance of their respective temporal grayscale sequences is calculated, resulting in 150 background variance values. The average of these background variance values ​​is then calculated. with standard deviation .like and The activity threshold For the aforementioned pixels The difference between its variance and the activity threshold is Pixel coordinates with a difference greater than zero will be marked. It was marked because its difference of 6.09 was greater than zero. An 8-neighborhood connectivity test was performed on all marked pixels, checking if the eight directly adjacent pixels of a marked pixel were also marked. If neighboring pixels were also marked, they were merged into the same region. This step was used to integrate discrete marked points into continuous regions. Finally, the pixel coordinates of these formed connected regions were collected to form a cytoplasmic candidate region coordinate set. This set contains several candidate regions, such as region A containing coordinates... Region B contains .

[0072] The membrane boundary extraction submodule is used to call the registration parameter set to transform the image coordinate system, extract the gradient sequence along the normal direction of the membrane structure edge, calculate the rate of change of gradient magnitude of adjacent pixels after normalization, detect abrupt change points where the rate of change exceeds the mean plus twice the standard deviation, record the coordinates of abrupt change points to form a spatial distribution set, and obtain the extreme points of the membrane boundary gradient.

[0073] The membrane boundary extraction submodule first calls the registration parameter set generated by the spatiotemporal registration module. This parameter set contains a sequence of affine transformation matrices for the transformation between each frame of images. These matrices are used to map the image coordinates of subsequent frames to the coordinate system of the reference frame (such as the first frame), ensuring that the cell membrane structure is aligned between frames. After alignment, based on the approximate outline of the cell (obtainable through preliminary segmentation), a certain number of pixels are taken along the estimated edge of the membrane structure, both inside and outside its normal direction, to form a grayscale value sequence. Consider a point P on the membrane, whose normal direction is... Along this normal direction, take 3 pixels inward and 3 pixels outward from point P, together with point P itself, to form a grayscale profile containing 7 pixels. ,in This is the grayscale value at point P. An actual sequence is... .

[0074] The first-order gradient magnitudes at these points are calculated using the central difference method, i.e. For point (Grayscale 85), its gradient magnitude is For point (Grayscale 105), its gradient magnitude is And so on, a gradient magnitude sequence is obtained. (The sequence length is 5, as bilateral differencing cannot be performed on edge points). Next, this gradient magnitude sequence is normalized using a minimum-maximum normalization method, mapping the gradient magnitudes to the [0,1] interval. For the sequence... Its minimum value maximum value Normalized gradient magnitude Substitute the values, Next, the rate of change of gradient magnitude between adjacent pixels after normalization is calculated, i.e. The rate of change sequence was obtained. Then, the rate of change sequence was detected. Points exceeding a specific statistical threshold are considered mutation points. The calculation steps for this statistical threshold are as follows: First, calculate the rate of change sequence. mean and standard deviation For sequences mean .variance Standard deviation Mutation detection threshold The rate of change sequence Each value in In comparison, if If so, the corresponding original pixel is marked as a mutation point. In this example, The maximum value in the sequence, 0.889, does not exceed 1.06437, indicating that the gradient change here does not reach the level of a sudden change. If in another set of data, a certain... If the value is 1.1, the point is marked. Record the spatial coordinates of all marked abrupt change points. These coordinates together form a spatial distribution set, and the points in this set are the extreme points of the membrane boundary gradient.

[0075] Table 1 Example of Inter-Frame Registration Parameters

[0076] Frame pairs Translation amount X (pixels) Translation amount Y (pixels) Rotation angle (degrees) Scaling ratio X Scaling ratio Y Spatiotemporal correlation weighting factor γ 1-2 3.10 2.40 1.10 1.015 0.995 0.938 2-3 -0.65 0.88 -0.29 0.997 1.004 0.961 3-4 1.25 -0.33 0.15 1.006 0.993 0.902 4-5 0.40 1.05 0.08 1.001 1.000 0.945 5-6 -0.95 -0.50 -0.18 0.994 1.007 0.913

[0077] As shown in Table 1, this table lists the translation difference (in pixels), rotation angle difference (in degrees), scaling difference, and the actual values ​​of the dynamically adjusted spatiotemporal correlation weight factor γ, calculated by the coordinate registration submodule between some adjacent frame pairs. These parameters constitute the registration parameter set and provide accurate inter-frame correspondences for subsequent image analysis.

[0078] In the weighted least squares method, the spatiotemporal correlation weight factor γ is dynamically adjusted according to the area ratio of the overlapping region in the image;

[0079] The normalization process uses the maximum-minimum standardization method to map the gradient magnitude to the [0,1] interval.

[0080] Please see Figure 3 The dynamic trajectory analysis module includes:

[0081] The optical flow tracing submodule is used to call the coordinate set of cytoplasmic candidate regions, establish the gray-level gradient matrix of pixels between adjacent frames, calculate the ratio of spatial gradient to temporal gradient in the optical flow equation, solve the displacement vector by the least squares method with regularization constraints, and perform median filtering on the vector using a three-frame sliding window to generate the initial trajectory vector set.

[0082] The optical flow tracing submodule first calls the set of coordinates of candidate cytoplasmic regions output by the cytoplasmic screening submodule, such as the set of pixels contained in candidate region A. The tracking process will focus on pixels within these regions. For any two adjacent frames, i.e., the registered first... frame and the frame Each pixel within candidate region A Choose one from the surrounding area A neighborhood window of a pixel is used to calculate the grayscale gradient information of pixels within that window, including the spatial gradient in the x-direction. Spatial gradient in the y-direction and time gradient . Specifically, , , For pixels In the Frames Neighborhood (for simplicity, actual usage) Its grayscale value is In its first The grayscale value at the corresponding position in the frame is 114. place, The value is calculated using the Sobel operator and is 3.5. The value calculated using the Sobel operator is 11.2; time gradient. .

[0083] Subsequently, the optical flow equations were solved using the least squares method with regularization constraints. displacement vector in The Lucas-Kanade method is used. For pixels... and Neighborhood Construct an overdetermined system of equations from these points. ,in , , The solution with Tikhonov regularization is The regularization parameter Based on the image signal-to-noise ratio setting, this value is set to 0.015. This parameter value is determined by performing optical flow calculation tests on a series of synthetic image sequences with different noise levels, selecting the value that minimizes the error between the calculated optical flow and the actual optical flow. Value. Solve for the displacement vector. (Units are pixels / frame). This process is applied to all pixels within the cytoplasmic candidate region to obtain the displacement vector field of each point between adjacent frames. Then, a sliding window containing tracking results from three consecutive frames is used to perform median filtering on these displacement vectors. For each pixel... At the point of time The calculated original displacement vectors are respectively , , After filtering Displacement vector at time The x-component is The median is 2.0; the y-component is... The median is -0.7. Therefore This filtering operation is performed on all tracking points and all time frames, ultimately generating an initial set of trajectory vectors, with each trajectory consisting of a series of filtered displacement vectors.

[0084] The orientation verification submodule is used to extract the displacement vectors of three consecutive frames of the initial trajectory vector set, calculate the cosine value of the angle between adjacent vectors, compare the cosine value with the orientation convergence threshold, mark the trajectory segments whose comparison results exceed the threshold for three consecutive times, and establish a candidate trajectory segment set.

[0085] The orientation verification submodule extracts trajectory segments from the initial trajectory vector set generated by the optical flow tracing submodule for analysis. Four consecutive frames of displacement vectors (all median filtered) are selected from a given trajectory: , , , Consider by Given the segments, calculate the cosine of the angle between adjacent vectors. The first angle... The cosine value is The second included angle The cosine value is .

[0086] The calculated cosine values ​​(1.000 and 0.9987) are numerically compared with a preset directional convergence threshold. This threshold is determined based on statistical analysis of a training dataset containing 300 known valid (i.e., representing real, continuous motion) intracellular particle trajectories. The cosine of the angle between each pair of adjacent displacement vectors within all these valid trajectory segments in the training set is calculated. The thousands of cosine values ​​obtained (which are typically very close to 1) are sorted, and their cumulative distribution function (CDF) is constructed. The cosine value corresponding to the 85th percentile of this CDF is selected as the directional convergence threshold. This method determines the directional convergence threshold. Next, we will examine the trajectory segment currently under investigation (composed of...). (Construction), check whether both of its calculated cosine values ​​are greater than That is, to judge. and In this case, both conditions are met. Therefore, the trajectory segment containing these three vectors is initially marked as having the same direction. If a trajectory involves vectors, as in this example... The cosine of the included angle is compared three times in a row (i.e. , , All results exceeded the threshold. The trajectory segment containing these four vectors is then marked as a candidate trajectory segment. By performing this direction verification process on all trajectory segments in the initial trajectory vector set, a candidate trajectory segment set is finally established, which contains all trajectory segments that have passed the initial direction consistency verification.

[0087] The trajectory filtering submodule is used to calculate the standard deviation of the displacement direction of multiple trajectory segments in the candidate trajectory segment set, calculate the difference between the standard deviation and the angle tolerance value, filter the trajectory segments with negative difference values, integrate the displacement vector parameters of the trajectory segments that meet the conditions, and obtain the effective trajectory vector set.

[0088] The trajectory filtering submodule processes each candidate trajectory segment in the candidate trajectory segment set output by the direction verification submodule. Consider a trajectory segment containing 12 consecutive displacement vectors. The candidate trajectory segments are identified. The standard deviation of the directions of all displacement vectors within each segment is calculated. Specifically, the standard deviation of each displacement vector is calculated first. Direction angle This yields a set of angle values ​​(expressed in radians). A set of actual angle values Calculate the arithmetic mean of this set of angle values. . Radius. Then calculate the standard deviation of the angle. Calculations yielded radians. This value is approximately equal to .

[0089] Next, calculate the standard deviation. The difference between the value and a preset angle tolerance value. Angle tolerance value This setting is based on long-term observation and data analysis of stable migration patterns of various cell types (including HUVEC and MDA-MB-231 cells) under different culture conditions. It represents the upper limit of acceptable directional fluctuation. radians (this value is approximately equal to) This value is set based on the fact that, after analyzing over 500 stable migrating cell trajectories, it was found that over 90% of these trajectories had an internal orientation angle standard deviation of less than [value missing]. The calculated difference is Radius. Then, filter out the differences. The trajectory segment with negative values, i.e. The trajectory segment. Because Therefore, this trajectory segment in the current example is preserved. For trajectory segments with positive differences (indicating significant directional fluctuations and a standard deviation greater than 1), the standard deviation is retained. If any of the selected short trajectory segments are not found to be valid, they are discarded. Finally, the displacement vector parameters of all the selected trajectory segments that meet the criteria are integrated. If multiple short trajectory segments that pass the screening can be continuously spliced ​​in time and space, they are merged into a longer valid trajectory. Through this series of operations, a set of valid trajectory vectors is obtained, which represents the high-confidence, directionally stable cell or particle motion trajectories in the system.

[0090] The directional convergence threshold is determined by the directional consistency distribution of the effective trajectory vectors in the statistical training set, and the 85th percentile value of the cumulative distribution function of this distribution is taken.

[0091] Please see Figure 4 The membrane surface deformation detection module includes:

[0092] The curvature gradient analysis submodule is used to obtain the extreme points of the membrane boundary gradient. It uses a curvature-area coupling model to simultaneously calculate the rate of curvature change and the rate of area change at the extreme points, establishes a curvature-area coupling parameter matrix, and obtains the extreme points of curvature gradient and the rate of area change through matrix eigenvalue decomposition.

[0093] The curvature gradient analysis submodule first obtains the set of extreme points with significant gradient changes on the membrane boundary, output by the membrane boundary extraction submodule. Within a specific image frame, points are identified on the cell membrane boundary... These extreme points are denoted as... Simultaneously, these points are captured in four consecutive frames (current frame). and the first three frames Location information within the range. For each extreme point... A curvature-area coupled model is used for analysis. This model simultaneously calculates the rate of curvature change and the rate of area change of the local region at a given point. First, the point... In the current frame Local curvature This curvature is obtained by fitting a quadratic curve to its neighborhood and then calculating it according to the formula; its value is... Simultaneously calculate a small area around it (with...). One of the centers The area within the template of a pixel (the region inside the cell). Its value is 28 square pixels. These values ​​are traced back to the previous frame. ,get for , It is 26 square pixels.

[0094] Rate of change of curvature .in This is the inter-frame time interval, set here to one unit of time (representing the actual physical time interval, such as 0.5 seconds, but represented as 1 frame in the calculation). Area change rate This calculation is performed on all 65 extreme points, resulting in 65 pairs. Values. Organize these parameter pairs into a... curvature-area coupling parameter matrix The first line is... Then, by examining this parameter matrix... covariance matrix (If the data is centralized, otherwise it is) (Subsequent processing) Eigenvalue decomposition is performed. This decomposition aims to identify the main coupling patterns of curvature and area changes. Eigenvectors indicate the main directions of change, and the corresponding eigenvalues ​​reflect the intensity of change in that direction. By analyzing these eigenvectors and eigenvalues, we can identify curvature gradient extrema that exhibit significant co-variance and obtain the specific curvature and area change rates corresponding to these points. We select the original extrema points that are primarily affected by the eigenvectors corresponding to the largest eigenvalue (representing the direction of greatest data variation), or points with large projection values ​​along this principal component direction, as key focus areas for subsequent deformation analysis.

[0095] The resistance index calculation submodule is used to call the curvature gradient extreme points and the area change rate, set the resistance coefficient weight in the curvature-area coupling model, perform a weighted superposition operation of the curvature change rate and the area change rate on the extreme points, and generate a resistance index distribution map corresponding to multiple extreme points.

[0096] The resistance index calculation submodule first calls the curvature gradient analysis submodule to analyze the rate of curvature change at a specific membrane boundary extreme point. and area change rate Data. For a selected extreme point ,That and In the curvature-area coupling model, resistance coefficient weights are assigned to these two rates of change, including the curvature change weight. and area change weight These weights are set based on the fact that, for deformation studies of specific cell types, drastic changes in curvature are often a better indicator of the cell membrane's mechanical response than small changes in local area; therefore, changes in curvature are given higher weights. and ,make sure These specific values ​​were determined through parameter tuning experiments on a series of cell samples known to exhibit specific biological processes, such as pseudopodia formation during the initiation phase of cell migration: in these samples, parameters were obtained from... arrive The resistance index was calculated using different weight combinations (step size 0.05), and then correlated with the actual membrane dynamics and mechanical response observed under a high-resolution microscope. The weight combination that best matched the observed phenomena was selected.

[0097] In addition, a cell membrane elastic modulus correction factor was introduced. The value of this factor is related to cell type; cell types with higher elastic modulus have higher elasticity. A higher value reflects a cell type's ability to resist deformation. For cell type A, with an elastic modulus of 2.0 kPa, For cell type B, with an elastic modulus of 4.5 kPa, . The specific range of values, such as 0.8 to 1.2, and their precise correspondence with cell types, were determined by measuring the Young's modulus of different cell lines using atomic force microscopy (AFM) and then performing regression analysis on these modulus values ​​and the degree of cell deformation under standard perturbation. The cells analyzed currently are of cell type A, therefore... Then, for this extreme point... The resistance index is calculated by weighting and superimposing the rate of change of curvature and the rate of change of area. Substitute the values: Repeating this calculation for all selected extreme points yields the resistance index corresponding to each extreme point. Correlating these resistance indices with their spatial locations on the membrane generates a resistance index distribution map corresponding to multiple extreme points. This map visually demonstrates the quantitative assessment of the cell membrane's resistance to deformation at different locations.

[0098] The deformation region location submodule is used to compare the resistance index gradient of adjacent extreme points based on the resistance index distribution map and the three-point extreme value detection method. It selects the extreme point group whose gradient change exceeds the set threshold, calculates the center point of the deformation region through three-point coordinate interpolation, and generates the deformation region coordinate set.

[0099] The deformation region localization submodule is used to process the resistance index distribution map corresponding to each membrane boundary extreme point generated by the resistance index calculation submodule. This map contains extreme points arranged sequentially along the membrane boundary. and its corresponding resistance index A three-point extreme value detection method is used to compare the gradient of the resistance index between adjacent extreme points. For any three consecutive extreme points... and its resistance index ,examine Does it constitute a local maximum or minimum? Simultaneously calculate the absolute values ​​of the gradients on both sides, i.e., the absolute value of the left gradient. and the absolute value of the right gradient Consider a sequence segment: Here It is a local maximum. , .

[0100] Select the group of extreme points where the gradient change exceeds a set threshold. This gradient change threshold is... Set as the absolute value of the local gradient of the resistance index on the current cell membrane (i.e., between all adjacent points). The maximum gradient value is 65% of the maximum gradient value. If the maximum absolute value of the difference in resistance index between all adjacent points on the membrane is calculated to be... Then the threshold This 65% percentage was obtained by retrospectively analyzing a large number of deformation events to determine the percentage of gradients that could effectively distinguish significant deformation regions from background fluctuations. In the example above, It did not exceed 0.195, but It exceeds 0.195. The filtering rule is: It is a local extremum, and and At least one of them is greater than Therefore, point group The selected set of three points is used to indicate the extreme points of deformation. For each selected set of three points... The coordinates of the center point of the deformation region are calculated by taking a weighted average of the coordinates of these three points. The weight is based on its resistance index value, i.e. , By collecting the coordinates of all the center points of the deformation regions obtained in this way, a set of coordinates of the deformation regions containing the center locations of multiple potential deformation regions is generated.

[0101] Table 2 Reference Values ​​for Cell Membrane Elastic Modulus Correction Factor η

[0102] Cell types Average Young's modulus (kPa) The setting value of the correction factor η Experimental verification samples (batch) Human umbilical vein endothelial cells (HUVECs) 1.8±0.4 0.92 15 Mouse fibroblasts (L929) 3.5±0.6 1.05 12 Human breast cancer cells (MCF-7) 0.9±0.3 0.85 18 Human cervical cancer cells (HeLa) 1.2±0.2 0.88 20

[0103] As shown in Table 2, this table lists the reference ranges of mean Young's modulus (measured by AFM) for several common cell types, as well as the correction factors set after calibration based on these moduli and cell mechanical behavior. The table provides specific numerical values, along with the number of experimental validation samples (or independent experimental batches) used to determine these values. This table is for the calculation of the resistance index. The selection of [the data] provides experimental basis and specific reference.

[0104] In the weighted superposition operation, a cell membrane elastic modulus correction factor η is introduced, and the value of η is positively correlated with cell type.

[0105] The gradient change threshold is set to 60%-70% of the maximum gradient of the resistance index.

[0106] Please see Figure 5 The subtype classification decision module includes:

[0107] The trajectory consistency analysis submodule is used to obtain the effective trajectory vector and the coordinates of the deformation area. It uses the support vector machine kernel function to map the trajectory direction feature space, calculates the vector dot product value between each trajectory point and the coordinate axis of the area, normalizes the dot product value to obtain the direction angle cosine sequence, and applies sliding window mean filtering to the cosine sequence according to the exponential decay time series weight to generate the trajectory direction consistency coefficient.

[0108] The trajectory consistency analysis submodule first acquires the effective trajectory vector set output by the dynamic trajectory analysis module (which contains the motion trajectories of several intracellular substructures or particles, each trajectory consisting of a series of spatiotemporal coordinate points and corresponding displacement vectors) and the deformation region coordinate set output by the membrane surface deformation detection module (which contains the center coordinates of regions with significant deformation activity, such as...). For a trajectory point on a specific trajectory. (Its in) The instantaneous velocity vector at time t is ) and a specified deformation region center First, a principal direction axis is defined for the deformation region. This principal direction axis is determined by analyzing the tensile direction of the deformation region or the average alignment direction of the stress fibers within the region. Let this direction vector be... Its numerical value is expressed as That is, the angle with the positive X-axis is 30 degrees. Then calculate the velocity vector of the trajectory point. With the coordinate axis vector of this region dot product value .set up ,but .

[0109] Next, the dot product value is normalized to obtain the cosine of the direction angle. For continuous trajectories These cosine values ​​of the frame A sliding window mean filter is applied using exponentially decaying time series weights. (Exponential decay factor) The value is set to 0.92, corresponding to the time window length. Set to 6 frames. These parameters ( This is based on the results of testing and optimization of a series of simulated trajectories (including orienteering, random walk, and mixed motion patterns). In different... (From 0.85 to 0.98, in steps of 0.01) and In a combination of frames (from 4 to 8), the calculated trajectory direction consistency coefficient is compared with the actual motion pattern of the trajectory, and the parameter combination that can optimally distinguish different motion patterns while also considering noise smoothing is selected. The weight sequence (unnormalized) is as follows: Normalized weights ,in The filtered mean If the cosine sequence of the past 6 frames is The calculated result That is, the trajectory point is in The consistency coefficient of the trajectory direction at any given time. Its value will be very close to the weighted average of the sequence, such as approximately 0.991.

[0110] The resistance mutation quantization submodule is used to divide the resistance mutation threshold interval based on the trajectory direction consistency coefficient using an adaptive Gaussian mixture model, calculate the second-order difference gradient value of the trajectory displacement in the X / Y axis of the deformation region, perform frame-by-frame difference operation on the absolute value of the gradient and the dynamic threshold, and calculate the cumulative amplitude sum of the points exceeding the threshold to generate the resistance mutation amplitude value.

[0111] The resistance mutation quantification submodule analyzes intracellular trajectories that interact with the deformed region. Its inputs include the trajectory orientation consistency coefficient calculated in the previous step (this coefficient is used to screen analysis objects; only trajectory segments with a consistency coefficient higher than 0.7 are analyzed, indicating a certain degree of motion persistence), and further examines the second-order dynamic characteristics of trajectory displacement. This involves analyzing the displacement sequence of an effective trajectory near the deformed region. (in (where the frame number represents a discrete time point), calculate its second-order difference gradient values ​​along the global X and Y axes. This value approximates the instantaneous acceleration. Second-order difference along the X axis. Second-order difference along the Y-axis The X-coordinate of a trajectory for three consecutive frames is... , , ,but Similarly, calculate... ,get Then calculate the second-order difference gradient in the current frame. amplitude .

[0112] Next, this gradient magnitude With a dynamic threshold Perform frame-by-frame differential calculations. This dynamic threshold... The update mechanism is .in It is the dynamic threshold (initial threshold) of the previous frame. The value is set based on the average acceleration fluctuation level of intracellular particles in the resting state; here it is 0.15. ), It is the average of the gradient magnitudes calculated for all tracked objects (or objects within a specific region) in the current frame. If The current frame is calculated as follows ,but Perform difference operations. Finally, all differences are statistically analyzed over the entire trajectory segment or a specific observation time window (e.g., 20 consecutive frames). Points greater than zero (i.e.) )of The cumulative sum of values If 6 consecutive frames Value (unit) ) are respectively Then the points exceeding the threshold The cumulative total is This value represents the magnitude of the generated resistance mutation, reflecting the cumulative fluctuation intensity of the trajectory beyond the dynamic baseline at the acceleration level.

[0113] The activation determination submodule is used to call the trajectory direction consistency coefficient and the resistance mutation amplitude value, establish a two-dimensional feature space coordinate system, construct a classification hyperplane using linear discriminant analysis, calculate the standardized Mahalanobis distance from the sample point to the decision boundary, divide the activation level interval according to the percentile of the distance distribution, and generate cell subtype activation labels.

[0114] The activation determination submodule first calls the trajectory direction consistency coefficient (e.g., generated by the trajectory consistency analysis submodule) ) and the resistance mutation magnitude value generated by the resistance mutation quantification submodule (e.g. To make it comparable to the consistency coefficient on a numerical scale, standardization or percentile transformation based on its distribution may be necessary. Here, we assume the calculated eigenvalues ​​are... These values ​​are either transformed or their original values ​​are used directly and processed by the LDA model. These two values ​​are used as the coordinates of a sample point in the two-dimensional feature space. Then, Linear Discriminant Analysis (LDA) is used to construct a classification hyperplane. The LDA is constructed based on a pre-labeled training dataset containing hundreds of cell samples with known biological activation states (such as "resting state," "early activation state," and "fully activated state") and their corresponding... Feature pairs. LDA finds an optimal projection direction (given by the weight vector) by maximizing the ratio of inter-class spread to intra-class spread. (Definition) This allows the projections of samples from different classes to be separated to the greatest extent possible along this direction. The classification hyperplane is perpendicular to this projection direction, and its equation can be expressed as: ,in , This is the bias term (threshold).

[0115] Next, calculate the current sample points to be classified. The standardized Mahalanobis distance to the decision boundary. The Mahalanobis distance calculation requires the pooled covariance matrix obtained from the training data. inverse matrix As the weight matrix. For a sample point and the mean vector of a certain category Its Mahalanobis distance to the center of the category is The distance to the LDA decision boundary is more directly calculated by projecting sample points onto the LDA discriminant axis, calculating the difference between the projected value and the classification threshold, and then standardizing it (usually by dividing by a measure of the within-class variance). The calculated standardized distance value is... Then, based on the distribution of standardized Mahalanobis distance values ​​calculated from a large number of samples, activation level intervals were defined using their percentiles. The criteria for defining activation level intervals are as follows: samples with standardized Mahalanobis distance values ​​between 0 and 0.7 are defined as "low activation"; samples between 0.7 and 1.8 are defined as "medium activation"; and samples greater than 1.8 are defined as "high activation". These thresholds (0.7 and 1.8) were determined through statistical analysis of the Mahalanobis distance distribution of various types of samples in the training set, such as using ROC curve analysis or equal-frequency partitioning methods, to achieve optimal classification performance. The current sample's... This value is greater than 1.8, therefore it falls into the range corresponding to "high activation". Finally, a corresponding cell subtype activation label, "high activation", is generated for this cell sample.

[0116] Table 3. Reference Table for Activation Level Range Division

[0117]

[0118] As shown in Table 3, this table provides a specific reference for dividing cell activation level intervals based on standardized Mahalanobis distance values. It not only lists the distance value range corresponding to each level and its percentile interval in the overall training sample distribution, but also gives the representative biological states that these levels may correspond to, providing a biological background for the interpretation of the results.

[0119] The exponential decay factor δ is 0.85-0.95, corresponding to a time window length of 5-7 frames;

[0120] The dynamic threshold update mechanism uses a weighted combination of 90% of the threshold of the previous frame and 10% of the average gradient magnitude of the current frame.

[0121] Mahalanobis distance calculation uses the inverse of the training set covariance matrix as the weight matrix.

[0122] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An immune cell image data classification and recognition system, characterized in that, The system includes: The spatiotemporal registration module is used to align the coordinate systems of multiple time-series images using an affine transformation algorithm based on a dynamic threshold of gray-level variance in spatiotemporal correlation. It filters candidate cytoplasmic regions based on a dynamically adjusted gray-level variance threshold, and transmits the coordinate set of candidate cytoplasmic regions to the dynamic trajectory analysis module. It also transmits the membrane boundary gradient extreme points to the membrane surface deformation detection module. The dynamic trajectory analysis module is used to call the optical flow equation parameter optimization method based on spatiotemporal constraints to analyze the coordinate set of the cytoplasmic candidate region, calculate the gray-level centroid displacement vector through a three-frame sliding window, verify the directional convergence using cosine similarity, and pass the effective trajectory vector to the subtype classification decision module. The membrane surface deformation detection module is used to calculate the resistance index of the membrane boundary gradient extreme point using a curvature-area coupling model that includes a curvature change rate weighting coefficient α and an area change rate weighting coefficient β. The module identifies the coordinates of the deformation region using a three-point extreme value detection method and transmits the coordinates of the deformation region to the subtype classification decision module. The dynamically adjusted grayscale variance threshold is determined based on the weighted average of the standard deviations of the grayscale distributions of three adjacent frames. The regularization parameter λ in the optical flow equation parameter optimization method has a value range of 0.1-0.3, which is used to balance the weight ratio of spatial gradient and temporal gradient. The curvature-area coupling model satisfies ,in Indicates the rate of change of curvature in the current frame. This represents the rate of change of the area in the current frame, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.

4. The subtype classification decision module is used to receive the effective trajectory vector and the coordinates of the deformation region, analyze the trajectory direction consistency coefficient using a support vector machine based on the radial basis kernel function, quantify the resistance mutation amplitude by a dynamic threshold segmentation algorithm combined with a sliding window standard deviation feedback mechanism, generate cell subtype activation labels, and output the cell subtype activation labels to the storage unit.

2. The immune cell image data classification and recognition system according to claim 1, characterized in that, The cytoplasmic candidate region coordinate set specifically includes the candidate region center coordinates and region boundary coordinates; the membrane boundary gradient extreme points include extreme point coordinates and gradient magnitude; the effective trajectory vector includes trajectory direction, displacement magnitude, and convergence coefficient; and the deformation region coordinates specifically include the deformation region center position and region boundary range.

3. The immune cell image data classification and recognition system according to claim 2, characterized in that, The spatiotemporal registration module includes: The coordinate registration submodule is used to acquire multi-frame temporal image data, establish a spatial transformation model between adjacent frames, calculate the difference in translation, rotation angle, and scaling ratio, linearly combine the difference parameters using an affine transformation matrix, iteratively optimize the matrix coefficients using the weighted least squares method, and generate a registration parameter set. The cytoplasmic screening submodule is used to call the registration parameter set to perform image coordinate transformation, extract the temporal grayscale sequence of multiple pixels, calculate the difference between the sequence variance value and the activity threshold, mark the pixel coordinates with a difference greater than zero in the continuous region, perform 8-neighborhood connectivity detection to merge fragmented regions, and establish a cytoplasmic candidate region coordinate set. The membrane boundary extraction submodule is used to call the registration parameter set to transform the image coordinate system, extract the gradient sequence along the normal direction of the membrane structure edge, calculate the rate of change of gradient magnitude of adjacent pixels after normalization, detect abrupt change points where the rate of change exceeds the mean plus twice the standard deviation, record the coordinates of abrupt change points to form a spatial distribution set, and obtain the extreme points of the membrane boundary gradient. In the weighted least squares method, the spatiotemporal correlation weight factor γ is dynamically adjusted according to the area ratio of the overlapping region of the image. The normalization process uses the maximum-minimum standardization method to map the gradient magnitude to the [0,1] interval.

4. The immune cell image data classification and recognition system according to claim 3, characterized in that, The dynamic trajectory analysis module includes: The optical flow tracing submodule is used to call the coordinate set of the cytoplasmic candidate region, establish the gray-level gradient matrix of adjacent frames, calculate the ratio of spatial gradient to temporal gradient in the optical flow equation, solve the displacement vector by the least squares method with regularization constraints, and perform median filtering on the vector using a three-frame sliding window to generate an initial trajectory vector set. The direction verification submodule is used to extract the displacement vectors of three consecutive frames of the initial trajectory vector set, calculate the cosine value of the angle between adjacent vectors, compare the cosine value with the direction convergence threshold, mark the trajectory segments whose comparison results exceed the threshold for three consecutive times, and establish a candidate trajectory segment set. The trajectory filtering submodule is used to statistically analyze the standard deviation of the displacement direction of multiple trajectory segments in the candidate trajectory segment set, calculate the difference between the standard deviation and the angle tolerance value, filter the trajectory segments with negative differences, integrate the displacement vector parameters of the trajectory segments that meet the conditions, and obtain the effective trajectory vector set. The directional convergence threshold is determined by the directional consistency distribution of the effective trajectory vectors in the statistical training set, and the 85th percentile value of the cumulative distribution function of this distribution is taken.

5. The immune cell image data classification and recognition system according to claim 4, characterized in that, The membrane surface deformation detection module includes: The curvature gradient analysis submodule is used to obtain the extreme points of the membrane boundary gradient. It uses a curvature-area coupling model to simultaneously calculate the rate of curvature change and the rate of area change of the extreme points, establishes a curvature-area coupling parameter matrix, and obtains the extreme points of curvature gradient and the rate of area change through matrix eigenvalue decomposition. The resistance index calculation submodule is used to call the curvature gradient extreme points and the area change rate, set the resistance coefficient weight in the curvature-area coupling model, perform a weighted superposition operation of the curvature change rate and the area change rate on the extreme points, and generate a resistance index distribution map corresponding to multiple extreme points. The deformation region positioning submodule is used to compare the resistance index gradient of adjacent extreme points based on the resistance index distribution map using the three-point extreme value detection method, select the extreme point group whose gradient change exceeds a set threshold, calculate the center point of the deformation region through three-point coordinate interpolation, and generate the deformation region coordinate set. The weighted superposition operation introduces a cell membrane elastic modulus correction factor η, and the value of η is positively correlated with the cell type. The gradient change threshold is set to 60%-70% of the maximum gradient of the resistance index.

6. The immune cell image data classification and recognition system according to claim 1, characterized in that, The cell subtype activation labels include trajectory direction consistency coefficient, resistance mutation magnitude, and subtype classification results.

7. The immune cell image data classification and recognition system according to claim 6, characterized in that, The initial threshold of the dynamic threshold segmentation algorithm is set as the moving average of the resistance mutation amplitude of the previous five frames.

8. The immune cell image data classification and recognition system according to claim 7, characterized in that, The subtype classification decision module includes: The trajectory consistency analysis submodule is used to obtain the effective trajectory vector and the coordinates of the deformation region, map the trajectory direction feature space using the support vector machine kernel function, calculate the vector dot product value between each trajectory point and the region coordinate axis, normalize the dot product value to obtain the direction angle cosine sequence, and apply sliding window mean filtering to the cosine sequence according to the exponential decay time series weight to generate the trajectory direction consistency coefficient. The resistance mutation quantization submodule is used to divide the resistance mutation threshold interval based on the trajectory direction consistency coefficient using an adaptive Gaussian mixture model, calculate the second-order difference gradient value of the trajectory displacement in the X / Y axis of the deformation region, perform frame-by-frame difference operation on the absolute value of the gradient and the dynamic threshold, and calculate the cumulative amplitude sum of the points exceeding the threshold to generate the resistance mutation amplitude value. The activation determination submodule is used to call the trajectory direction consistency coefficient and the resistance mutation amplitude value to establish a two-dimensional feature space coordinate system, construct a classification hyperplane using linear discriminant analysis, calculate the standardized Mahalanobis distance from the sample point to the decision boundary, divide the activation level interval according to the percentile of the distance distribution, and generate cell subtype activation labels. The exponential decay factor δ is 0.85-0.95, corresponding to a time window length of 5-7 frames; The dynamic threshold update mechanism uses a weighted combination of 90% of the threshold of the previous frame and 10% of the average gradient magnitude of the current frame. The Mahalanobis distance is calculated using the inverse of the training set covariance matrix as the weight matrix.

Images